Network based search services, Internet search engines, voice search, local search, and various other technologies for searching and retrieving information have become increasingly important for helping people find information. Voice search involves a coupling of voice recognition and information retrieval. An uttered phrase is automatically recognized as text, and the text is submitted as a query to a search service. For example, a person may use a mobile phone equipped with a voice search application to find a restaurant by speaking the name of the restaurant into the mobile device, and the mobile device may recognize the spoken restaurant name (i.e., convert it to text) and transmit the text of the restaurant name to a remote search service such as a business directory. Local search is a special case of search where listings of business establishments, firms, organizations, or other entities have been used to enable mobile devices to search same. Consider the following example.

A user may be interested in finding information about a business listed in a directory as "Kung Ho Cuisine of China". However, the user formulates a query as "Kung Ho Restaurant". Currently, a search for this listing will not take advantage of statistical parallels between parts of the query and listing forms. Furthermore, erroneous listings, e.g. "Kung Ho Grocery" may be returned as a relevant match.

Discussed below are techniques related to statistical intra-language machine translation, and applications thereof to speech recognition, search, and other technologies.

Stefan Riezler et al.: "Statistical Machine Translation for Query Expansion in Answer Retrieval", ALC 2007 Proceedings, June 23, 2007, relates to system and methods for query expansion in answer retrieval that uses Statistical Machine Translation techniques to bridge the lexical gap between questions and answers. Statistical Machine Translation (SMT)-based query expansion is done by i) using a full-sentence paraphraser to introduce synonyms in context of the entire query, and ii) by translating query terms into answer terms using a full-sentence SMT model trained on question-answer pairs. We evaluate these global, context-aware query expansion techniques on tfidf retrieval from 10 million question-answer pairs extracted from FAQ pages. Experimental results show that SMTbased expansion improves retrieval performance over local expansion and over retrieval without expansion.

Adam Berger et al.: "Information Retrieval as Statistical Translation", PROC. SIGIR 1999 (Annual International ACM-SIGIR conference on research and development in information retrieval), University of California, Berkeley, USA, August 15, 1999, pages 222-229, describes a new probabilistic approach to information retrieval based upon the ideas and methods of statistical machine translation. The central ingredient in this approach is a statistical model of how a user might distill or "translate" a given document into a query. To assess the relevance of a document to a user's query, the probability that the query would have been generated as a translation of the document is estimated, and factor in the user's general preferences in the form of a prior distribution over documents. A simple, well motivated model of the document-to-query translation process is proposed, and it is also described an algorithm for learning the parameters of this model in an unsupervised manner from a collection of documents.

C. Quirk et al.: "Monolingual Machine Translation for Paraphrase Generation", Proceedings of the Conference on Empirical Methods in Natural Language Processing, July 25, 2004, pages 142-149. In this document, statistical machine translation tools are applied to generate novel paraphrases of input sentences in the same language. The system is trained on large volumes of sentence pairs automatically extracted from clustered news articles available on the World Wide Web. Alignment Error Rate is measured to gauge the quality of the resulting corpus. A monotone phrasal decoder generates contextual replacements. Human evaluation shows that this system outperforms baseline paraphrase generation techniques and, in a departure from previous work, offers better coverage and scalability than the current best-of-breed paraphrasing approaches.

Josep M. Crego, Adria de Gispert, Jose B. Marino.: "The TALP NGRAM-Based SMT System for IWSLT'05" October 24, Pittsburgh, PA, USA, provides a description of TALP-Ngram, the tuple-based statistical machine translation system developed at the TALP Research Center of the UPC (Universitat Politecnica de Catalunya). Briefly, the system performs a log-linear combination of a translation model and additional feature functions. The translation model is estimated as an N-gram of bilingual units called tuples, and the feature functions include a target language model, a word penalty, and lexical features, depending on the language pair and task.

It is the object of the present invention to provide a system and a method for automatically translating between spoken queries and listings.

This object is solved by the subject matter of the independent claims.

Preferred embodiments are defined by the dependent claims.

The following summary is included only to introduce some concepts discussed in the Detailed Description below. This summary is not comprehensive and is not intended to delineate the scope of the claimed subject matter, which is set forth by the claims presented at the end.

Training data may be provided. The training data may include pairs of source phrases and target phrases. The pairs may be used to train an intra-language statistical machine translation model, where the intra-language statistical machine translation model, when given an input phrase of text in the human language, can compute probabilities of semantic equivalence of the input phrase to possible translations of the input phrase in the human language. The statistical machine translation model may be used to translate between queries and listings. The queries may be text strings in the human language submitted to a search engine. The listing strings may be text strings of formal names of real world entities that are to be searched by the search engine to find matches for the query strings.

Many of the attendant features will be explained below with reference to the following detailed description considered in connection with the accompanying drawings.

The present description will be better understood from the following detailed description read in light of the accompanying drawings, wherein like reference numerals are used to designate like parts in the accompanying description.

• Figure 1 shows a general process for intra-language statistical machine translation.
• Figure 2 shows a process for building an n-gram based model.
• Figure 3 shows an arrangement for using a statistical translation model to improve a search system and/or the language model of a voice recognition system.

The description below covers embodiments related to using a statistical machine translation model to translate between sentences or phrases of a same human language. The description begins with discussion of how a relatively small set of training sentences or phrases are used to train a statistical translation model. Applications of the intra-language machine translation model are then described, including applications to search, automatic speech recognition (ASR), and display of speech recognition results

Statistical models have been used to translate sentences from one language to another language. However, they have not been trained or used for translating between phrases or sentences of a same language. That is, statistical modeling has not previously been used to translate phrases in English, for example, to other semantically similar phrases also in English.

A statistical translation model is a generalization of some sample of text, which may be parallel phrases such as a query strings and corresponding directory listings. Some types of statistical translation models give probabilities that a target sentence or phrase is a translation of a source sentence or phrase, and the probabilities reflect the statistical patterns derived from the training text. In effect, the model is a probabilistic generalization of characteristics or trends reflected from statistical measurements of training sentences. Note that throughout this description, the terms "sentence" and "phrase" will be used interchangeably to refer to relatively short arrangements of words. Formal and informal names of businesses, query strings inputted by users, grammatical sentences, clauses, and the like are examples of sentences or phrases. Note also that while this description discusses intra-language statistical machine translation as applied to phrase-based search (in particular voice and/or geographically localized search), the concepts are not limited to search applications. Furthermore, searching listings of short phrases is also applicable to other types of search besides local search, including product search, job search, etc.

Figure 1 shows a general process for intra-language statistical machine translation. Initially, a statistical machine translation model is trained 100. Training 100 will be described in detail later. The training 100 is performed using a sample of training data, which may come from a variety of sources. The training data will include parallel (paired) phrases in a same human language. The training 100 informs the translation model with statistics (e.g., n-grams) that can be used to compute probabilities or likelihoods of candidate translations of a phrase. Specific training 100 for an n-gram based model will be described below.

After the model is trained 100, the model is used to translate 102 a source phrase into a target phrase. Translation 102 involves starting with a source phrase and obtaining a semantically similar or equivalent target phrase. For example, a source phrase "Kung Ho Cuisine of China" might be translated to a target phrase "Kung Ho Chinese Restaurant" or "Kung Ho Restaurant" Different forms of candidate target phrases are obtained. The statistical translation model is used to find one or more of the most probable candidate target phrases. Consider the following overview of voice based search and how it relates to intra-language machine translation.

A voice search system may involve two components: a voice recognition component and an information retrieval (search) component. A spoken utterance o is converted into a text query q using automatic speech recognition (ASR), i.e., $$q^{*}={\mathrm{argmax}}_{q}p\left(o|q\right)p\left(q\right)$$ where p(o|q) and p(q) represent an acoustic model and a language model (LM), respectively. Statistical LMs, e.g. n-gram models, are often used to allow flexibility in what users can say. That is, they allow a variety of sayings to be recognized by the ASR component. Next, the best (or n-best) q is passed to a search engine to retrieve the most relevant document d, i.e. $$d*={\mathrm{argmax}}_{d}p\left(d|q\right)$$

In the context of local search, documents d may be in the form of business listings (names of business, organizations, or other entities), which are typically short, e.g. "Kung Ho Cuisine of China".

Given this framework for voice based search, listings and queries, because they are both relatively short, are treated as pairs akin to "sentence pairs" found in bilingual translation training. A bilingual statistical translation model, adapted for intra-language translation, may be used to automatically convert the original form of a listing to its query forms (i.e., the forms that a user might be expected to input when searching for the listing), which in turn may be used for building more robust LMs for voice search, grammar checking, or other applications. Conveniently, the statistical translation model may be trained using a small number of transcribed or artificially produced queries, without necessarily having to acquire matching listings. While a variety of types of statistical models can be used for machine translation, an n-gram based model will be described next.

Although a query phrase and its intended listing phrase may differ in form, there is usually a semantic correspondence, at the word level, between the two phrases. In other words, words in the query can be mapped to words in the listing or to a null word, and vice versa. A machine translation approach may be used to predict possible query forms of a listing, and then to utilize the predicted query forms to improve language modeling. Specifically, as discussed next, n-grams may be used on word pairs to model the joint (conditional) probability of a listing and a query.

Figure 2 shows a process for building an n-gram based model. A pair of source and target sentences in a same human language are received 120. An alignment between the source and target sentences is obtained 122 by computing an edit distance between the two sentences. Words and/or phrases of the aligned sentences are then paired 124 and treated as semantic units. Pairings may be formed by finding semantically/literally similar/equivalent words or phrases. The pairings are then used to train 126 an n-gram model. The steps of this process may be repeated for different source and target sentences. While a small set of training sentences may suffice for some applications, using more training data will create a more robust model. Note also that the alignment and the n-gram model may be iteratively updated and refined in the maximum likelihood sense.

Details of generating an n-gram based model will now be described. For training 100 an n-gram based model, initial training data is provided. This data may be a body of parallel text (d, q), where listings d and queries q serve as source and target sentences respectively. The sentences d and q may be monotonically aligned, where null words are added, if necessary, to account for insertions or deletions that occur in the alignment. The monotonic alignment will be denoted as a. Note that in another embodiment, a non-monotonic alignment may be used.

Once aligned, a sequence pairs of words from d and q is generated, which is denoted as (d, q, a) = ((d₁, q₁), (d₂, q₂), ..., (d_L, q_L)), where each (d_i, q_i) is treated as a single semantic unit. Consecutive word pairs can be merged to form phrase pairs if necessary.

The sequence of word pairs can then be used to train an n-gram model. Consequently, the probability of an aligned sentence pair is computed as $$\begin{array}{l}\left(d,q,a\right)=\\ {\displaystyle {\prod }_i^{p_M}p\left(\left(d_i,q_i\right)|\left(d_{i-n+1},q_{i-n+1}\right),\dots ,\left(d_{i-1},q_{i-1}\right)\right)}\end{array}$$ where M denotes the monotonic condition. Note that the initial alignment a may be computed using the Levenshtein distance between d and q. The alignment and the n-gram model's parameters may be updated in the maximum likelihood sense. Re-alignment can be based on pairing frequencies, for example.

Given the trained n-gram model, a listing-to-query translation may be performed. Given a listing form d, and given query forms q (from a decoder, discussed later), the query forms are searched to find those that have the highest conditional probability: $$q^{*}\approx {\max }_q{\max }_a{p}_M\left(d,q,a\right)$$ where p(d, q, a) is evaluated using equation (3).

The translation not only exploits word-level semantic correspondence as modeled by unigrams, but it also takes into account word context by using higher order n-grams. The search for the best or n-best query forms can be achieved efficiently by applying the best-first search algorithm, which is described by Russell and Norvig in Artificial Intelligence: A Modern Approach (Prentice Hall, second edition, 2003). Using this type of search, pruning techniques may be applied to reduce computational complexity. Returning to the language model (LM) for speech recognition, once the n-best query forms are obtained for the listings, they may be used as training sentences for LM estimation.

There are two implementation details to be considered. First, allowing the use of null words in d raises a potential problem at decode time -- the search space is significantly expanded because null can be present or absent at any position of the source sentence. To avoid this problem, it is preferable to eliminate the use of (d_i = null, q_i) as semantic units for values of q_i. Specifically, in training, (d_i = null, q_i) may be merged with its proceeding or following semantic unit, depending on which of the phrases, q_i-1q_i or q_iq_i+1, have more occurrences in the training data. Then, (d_i+1, q_i-1q_i) or (d_i+1, q_iq_i+1) may be treated as a single semantic unit. At decode time, null is not explicitly inserted in d, because using semantic units (d_i-1, q_i-1q_i) or (d_i+1, q_iq_i+1) is equivalent to adding null in the source sentence.

The second implementation detail concerns out-of-vocabulary (OOV) words in d. When OOV occurs, it might not be possible to produce any query forms, since p(d_i = OOV, q_i) = 0 for any value of q_i. To deal with such cases, a positive probability may be assigned to unigrams (d_i , q_i = d_i) whenever d_i = OOV. This implies that a listing word, if never seen in the training data, will be translated to itself.

It should be noted that embodiments with non-monotonic alignment are also possible. Furthermore, a re-ordering strategy may be used. This may be implemented before a monotonic alignment is applied by reordering d while keeping the order of q. When training the translation model, the best way to reorder the words in the source form is determined by computing the resulting joint n-gram model likelihood. Only orders that are shifts of the original order are considered, and a maximum entropy classifier for these orders is built, where the input of the classifier is the source form, and the output is an order. Prior to translation, this classifier is applied to reorder a source form.

Figure 3 shows an arrangement for using a statistical translation model to improve a search system and/or the language model of a voice recognition system. A search engine 152 is configured to search listings 154, for example business listings. The search engine 152 receives text queries or transcribed spoken queries 156 that are generated by users and submitted to the search engine 152. Corresponding relevant listings 158 are retrieved by the search engine 152. Note that training pairs can also be obtained algorithmically using TF-IDF (term frequency-inverse document frequency).

The text or transcribed queries 156 and corresponding search-engine retrieved listings 158 are passed to a training component 160 that trains a statistical translation model 162, which may be n-gram based or another type of model. As discussed above, the training component 160 iterates through source-target pairs of the transcribed queries 156 and listings 158. In the case of an n-gram based model, given a (source, target) pair, an initial monotonic alignment is obtained between the source form and target form by computing an edit distance. Given the alignment, the training component 160 discovers word-level pairs and builds an n-gram translation model 162 based on the word-level pairs. The alignment and n-gram model parameters of the translation model 162 may be iteratively refined to improve the translation model 162. Furthermore, training may implement a backoff strategy which assumes that a word can be translated to itself, as is possible with intra-language translation. In other words, the aligned units WORD-WORD, where WORD can be a word or a phrase, will have a positive probability.

A translation module 164 uses the translation model 162 to test decoded candidates (potential translations). Given the trained translation model 162 and a source form, a best-first search algorithm may be used to obtain the top n-best target forms (the n decoded target forms with the highest probability according to the translation model 162). The weight of each target form is determined by p(target|source) produced by the translation model. Unlikely word-level pairs may be pruned to speed up translation.

Given the translation model 162 and the translation module 164, subsequent searches may be improved as follows. Given a user's query q and a listing d found by the search engine 152, translated query forms x of the listing d are considered when measuring the listing d's relevancy to a user's query. Letting s(_,_) be a function or measure of relevancy (or similarity), the measurement of relevancy may be s(q, d) = sum_x {p(x|d) s(q, x)}. Alternatively, relevancy may be measured directly from the translation probability, in which case s(q, d) = p(q,d). In one embodiment, potential translations can be filtered out if their similarity measure is below a specified threshold.

Furthermore, not only may searching be improved as described above, but a language model 168 can also be built or augmented using intra-language translation. Language models are used in many natural language processing applications such as ASR, machine translation, and parsing. The intra-language translation provided by the translation model 162 and translation module 164 may be used in language modeling by translating listings into query forms and using the same-language translated query forms when estimating a language model 168. When estimating the language model 168, the count of a translated query form may be set to its posterior probabilities multiplied by the count of its original listing.

In one embodiment, a server- or client-based voice recognizer may be provided with the language model 168, which will allow the voice recognizer to perform more accurate and comprehensive speech recognition with respect to utterances directed to the listings 154 or to listings. The translation model 162 may also be used at a server or at a mobile client to translate a string inputted at the mobile device (whether by ASR or otherwise) to a display form.

Embodiments and features discussed above can be realized in the form of information stored in volatile or non-volatile computer or device readable media. This is deemed to include at least media such as optical storage (e.g., CD-ROM), magnetic media, flash ROM, or any current or future means of storing digital information. The stored information can be in the form of machine executable instructions (e.g., compiled executable binary code), source code, bytecode, or any other information that can be used to enable or configure computing devices to perform the various embodiments discussed above. This is also deemed to include at least volatile memory such as RAM and/or virtual memory storing information such as CPU instructions during execution of a program carrying out an embodiment, as well as non-volatile media storing information that allows a program or executable to be loaded and executed. The embodiments and featured can be performed on any type of computing device, including portable devices, workstations, servers, mobile wireless devices, and so on. The modules, components, processes, and search engine 152 discussed above may by realized on one computing device or multiple cooperating computing devices.